The Microtargeting Claim: Precision without proof

In brief

A political message written for you personally persuades you no more than a good message written for everyone. Pooled across 76,977 participants, the gain from tailoring was 0.43 of a percentage point, and the one clean head-to-head test put individually written messages slightly behind a single universal one. What does look effective is choosing which message to show a group in the first place, which is ordinary message testing rather than personalisation of wording. Two caveats keep this from being final: nearly all the evidence comes from survey experiments that measure stated issue positions, not votes, and that design tends to understate real effects rather than flatter them.

How to use this

Ask which of the two jobs a targeting product actually does: picking which message goes to whom, or rewriting it for each person. Test one good general message against a handful of alternatives first, because that is where the documented gains sit. Ask for evidence that votes in a real election have moved, not just stated issue positions; if nobody can show it, name the gap rather than accept a figure. When you meet a striking claim about personalisation, check what it was compared against. Treat data volume as no promise of effect: profiles built from one piece of information did as well as those built from four.

What the story is about

The cleanest test of tailoring a message to one person took one person's own answers about their politics, fed them to a language model, and had it write a message for that person alone. A pre-registered experiment, meaning its plan was fixed in advance, ran that on 8,587 people across four political issues (Hackenburg and Margetts, 2024a). Persuasion here is measured in percentage points, meaning a shift in the share of people who agree with a position. One point is one person in a hundred. The individually written messages moved support by 4.83 points. A single message written for everybody moved it by 6.20 points. So tailoring did not beat the general message. It came out slightly behind, and the gap was small enough that chance is a plausible explanation (P = 0.226).

Single experiments can mislead, so the largest test pooled its own three. The largest exercise of its kind ran three experiments with 76,977 participants, 19 language models and 707 political issues (Hackenburg et al., 2025). Pooled across every method and every study, the authors report a gain over a non-personalised message of 0.43 of a percentage point, with the plausible range running from 0.22 to 0.64. No single method reached one point. More data did not help: profiles built from one, two and four pieces of information performed alike, and the authors conclude this approach "may not require the collection of vast amounts of personal data" (Hackenburg et al., 2025). What did predict persuasion was information density, the number of checkable claims a message carried. Where persuasiveness rose, factual accuracy fell. Two smaller studies agree. Matching a message to an issue position someone already holds works; matching its phrasing to their personality does not (Decker and Krämer, 2023; Walker, O'Neill and Wit, 2020).

Targeting covers two jobs: choosing which message to show someone, and writing that message differently for them. The famous advantage comes almost entirely from the first. Tappin et al. (2023) used a model to choose which ad went to which group, then tested it against one best message for everyone. Microtargeting won on the pooled result, 5.96 points to 3.48 (P = 0.004). Split by issue, the whole win came from one of the two. On the other, the scores were 7.16 and 5.03, and the difference was not reliable. When the same method chose which attitude to target rather than which message, it scored 12.30 points against 15.42 for the single best message, which is a loss. So the widely quoted figure of 70% or more rests on a comparison with one message. In the 2022 Danish general election, parties used online targeting to reach people who already voted for them (Hove et al., 2026).

Other researchers pushed back in print, and the exchange repays reading because neither side claims what the headlines say. Teeny and Matz (2024) argued that finding no effect does not prove these models cannot personalise a message. They offered two explanations: the models may not have personalised enough, and the ten simple attributes used, such as age and party, may be too thin next to richer psychological detail. Hackenburg and Margetts (2024b) accepted the first point, pointing to a marginal but statistically reliable sign of personalisation in their own appendix. They rejected the second, since party, race, age and gender are the attributes political science has found shape how people respond to messages. Their conclusion is narrow: personalising with today's models is not guaranteed to raise persuasiveness.

Nearly every number so far comes from a survey experiment rather than a campaign. People are recruited, shown a message, and asked what they think straight away (Hackenburg and Margetts, 2024a; Tappin et al., 2023). That setup strips out the things that make real advertising hard to land: getting attention, competing with everything else in a feed, and the memory fading within days. Those forces push measured effects down, which means the estimates here are more likely to understate a real campaign than to flatter it. What moves is also narrower than a vote. The outcome measured is a stated position on an issue. No study in this record estimates how microtargeting changes votes in a real election.

A message fitted to what a campaign knows about you persuades you about as much as a good general one. The measured premium for tailoring comes in below half a percentage point, and the one clean head-to-head test put it slightly below zero. Where personalisation seemed to reach further, the gain came from choosing a better message for a group rather than rewriting one for a person, which is message testing. The direction of travel is the same across every design in this record: match the message to the position someone already holds, and spend your effort there.

So what

What moves people is the substance of the message and how well it fits what they already believe. Persuasion tracked how much checkable substance a message carried, and messages landed hardest when they matched a position someone already held. Matching on what someone already thinks is the part that worked. One good message, tested against a few others, beats a different message for every voter. More personal data bought no extra movement.

For political parties

A party would most want to know whether targeting changes votes in a real campaign. That gap, noted above, is the one a party would most need closed. The conversion payoff is therefore unproven, and any party buying a targeting system is buying an unmeasured claim. What is documented sits on the mobilisation side: parties aim online targeting at people who already vote for them. So the honest question is which number the spend is meant to move, and whether anyone has shown that number moving. Ask it before signing, not after.

For government

What a government would like is a price tag: so much spent on targeting, so many people moved. Nobody has produced one. Reach is priced one way and persuasion is measured another, and no study in this record joins the two under different targeting strategies. Anyone adding them up is doing their own arithmetic, from two sources that were never connected. So nobody can say in money what targeting is worth, which is awkward for a government writing rules about it. A rule is being written ahead of the arithmetic, and the honest thing is to say so: regulate collection and transparency, and treat claims about persuasive payoff as unproven rather than settled. Where the evidence is thin, name the gap instead of filling it with a number that sounds authoritative.

Case studies

The case everyone cites for what microtargeting can do is Cambridge Analytica. The primary record is the Information Commissioner's Office investigation, which began in 2017, seized and analysed the company's servers, and reported to Parliament in October 2020 (Information Commissioner's Office, 2020). That report is a regulator's account of what it found, not peer-reviewed research. It describes bulk purchased commercial data, "at one estimate over 130 billion data points", and models built mainly from off-the-shelf tools. It also records that "there was evidence that their own staff were concerned about some of the public statements the leadership of the company were making about their impact and influence". Ordinary data, ordinary tools, and a leadership making larger claims than the work supported: that is what the case documents.

References

Decker, H. and Krämer, N. (2023) 'Is personality key? Persuasive effects of prior attitudes and personality in political microtargeting', Media and Communication, 11(3), pp. 250-261. Available at: https://doi.org/10.17645/mac.v11i3.6627 (Accessed: 10 September 2026).

Hackenburg, K. and Margetts, H. (2024a) 'Evaluating the persuasive influence of political microtargeting with large language models', Proceedings of the National Academy of Sciences, 121(24)., e2403116121. Available at: https://doi.org/10.1073/pnas.2403116121 (Accessed: 10 September 2026).

Hackenburg, K. and Margetts, H. (2024b) 'Reply to Teeny and Matz: toward the robust measurement of personalized persuasion with generative AI', Proceedings of the National Academy of Sciences, 121(43)., e2418817121. Available at: https://doi.org/10.1073/pnas.2418817121 (Accessed: 10 September 2026).

Hackenburg, K. et al. (2025) 'The levers of political persuasion with conversational artificial intelligence', Science, 390(6777)., eaea3884. Available at: https://doi.org/10.1126/science.aea3884 (Accessed: 10 September 2026).

Hove, M.F. et al. (2026) 'Preaching to the choirs: political parties' online targeting strategies in multi-party systems', Party Politics. Advance online publication. Available at: https://doi.org/10.1177/13540688251408777 (Accessed: 10 September 2026).

Information Commissioner's Office (2020) ICO investigation into use of personal information and political influence. Letter ICO/O/ED/L/RTL/0181 from Elizabeth Denham, UK Information Commissioner, to Julian Knight MP, Chair of the Digital, Culture, Media and Sport Committee, 2 October 2020. Available at: https://committees.parliament.uk/publications/2847/documents/27859/default/ (Accessed: 10 September 2026).

Tappin, B.M. et al. (2023) 'Quantifying the potential persuasive returns to political microtargeting', Proceedings of the National Academy of Sciences, 120(25)., e2216261120. Available at: https://doi.org/10.1073/pnas.2216261120 (Accessed: 10 September 2026).

Teeny, J.D. and Matz, S.C. (2024) 'We need to understand "when" not "if" generative AI can enhance personalized persuasion', Proceedings of the National Academy of Sciences, 121(43)., e2418005121. Available at: https://doi.org/10.1073/pnas.2418005121 (Accessed: 10 September 2026).

Walker, C., O'Neill, S. and de-Wit, L. (2020) 'Evidence of psychological targeting but not psychological tailoring in political persuasion around Brexit', Experimental Results, 1., e38. Available at: https://doi.org/10.1017/exp.2020.43 (Accessed: 10 September 2026).